SEARCH · Search NASA

Results for “neural function”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

SourceResearch connectionFrom yearThrough year

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 217 records · Page 12

All sources NASA NTRS DOE OSTI

DOE OSTI2025

Triangle Method for Dense ReLU Layers [SWR-25-72]

This software is an implementation of the methods for initializing and training neural networks to be more efficient per parameter, described more fully below and in the related publication: In theory, depth should make a ReLU network EXPONENTIALLY more efficient by enabling it to produce an exponential number of piecewise linear sections in its output. This reasoning is largely based on the work of mathematicians that have hand-constructed networks that make good use of depth. In practice however, even very deep ReLU networks that have been randomly initialized will behave identically to their shallow counterparts - missing an entire exponential dimension of efficiency. The triangle method is a first attempt at realizing the exponential potential of deep networks. Instead of randomly setting weights, we force pairs of neurons in each layer learn to build triangles (i.e. functions from [0,1] -> [0,1] that look like triangles). This is a very efficient pattern for generating lots of linear pieces because composing two triangular functions doubles the number of pieces with each composition. The triangle method is more than just a different initialization, it is a new paradigm of training. Instead of making direct updates to the matrix weights, we do an extra step of backpropagation to collect the derivatives of the loss function with respect to the shapes of the triangles, training them to tilt left or right. This process essentially holds the networks hand throughout the loss landscape and forces it to always use depth effectively by producing triangular shapes internally. This can produce several orders of magnitude of improvement on convex one-dimensional regression problems. Much more theoretical work is needed to realize its full potential beyond this context, but the implementation in this repository will still work in arbitrary numbers of dimensions. The file Triangle_Method.py is a generalized form of the method that will build each neuron its own custom 1-d convex activation function (with exponential efficiency). Example usage on one dimensional problems can be found in Example_Usage.ipynb and an example of using this in a real neural network can be found in Example_VGG16_CIFAR10.ipynb.

Milkert, Max [National Renewable Energy Laboratory↗

DOE OSTI2025

Femtojoule optical nonlinearity for deep learning with incoherent illumination

Optical neural networks (ONNs) are a promising computational alternative for deep learning due to their inherent massive parallelism for linear operations. However, the development of energy-efficient and highly parallel optical nonlinearities, a critical component in ONNs, remains an outstanding challenge. Here, we introduce a nonlinear optical microdevice array (NOMA) compatible with incoherent illumination by integrating the liquid crystal cell with silicon photodiodes at the single-pixel level. We fabricate NOMA with more than half a million pixels, each functioning as an optical analog of the rectified linear unit at ultralow switching energy down to 100 femtojoules per pixel. With NOMA, we demonstrate an optical multilayer neural network. Our work holds promise for large-scale and low-power deep ONNs, computer vision, and real-time optical image processing.

36 MATERIALS SCIENCE↗

DOE OSTI2025

Ensemble cure kinetics network (ECK-Net): A method to derive cure kinetics of thermosetting resin

This paper introduces an Ensemble Cure Kinetics Network (ECK-Net), a neural network (NN)–based framework for modeling the cure kinetics of thermosetting resins within a phenomenological context. ECK-Net replaces traditional analytic models, which require extensive chemical insight and multiple isothermal/non-isothermal experiments, with a data-driven surrogate that maps nonlinear relationships between temperature, degree of cure, and reaction rate from differential scanning calorimetry data. The proposed approach predicts input-dependent kinetic coefficients of a generalized nth-order reaction equation rather than reaction rates directly, enabling a single unified model to represent various epoxy systems without relying on iso-conversional analysis or predefined functional forms. To ensure robustness, multiple independently trained networks under different random initializations are blended through an ensemble strategy, effectively mitigating the stochastic variability inherent to neural networks. The framework is validated using experimental datasets from multiple resin systems, including aerospace-grade materials (Toray 3900-2, Cycom 5320-1, and Hexcel 8552) and a windmill-grade resin (RIMR 035c). The model accurately reproduces the temporal evolution of the degree of cure under manufacturers’ recommended cure cycles across all tested resins systems, yielding Pearson’s correlation coefficients of 0.992, 0.994, 0.993, 0.997, respectively. To demonstrate process-level applicability, the trained network was implemented within the Abaqus environment to simulate out-of-autoclave (OOA) curing process of the CFRP panel composed of Toray T830H-6K/3900-2D prepreg. The simulation results showed excellent agreement with experimental temperature response (maximum peak temperature, simulation: 189.6 °C, experiment: 188.5 °C) and the final degree of cure (simulation: 0.948, experiment: 0.960 ± 0.013), confirming ECK-Net’s capability as a reliable alternative to conventional cure kinetics modeling methods.

Composite curing↗

DOE OSTI2025

CryoDRGN-AI: neural ab initio reconstruction of challenging cryo-EM and cryo-ET datasets

Proteins and other biomolecules form dynamic macromolecular machines that are tightly orchestrated to move, bind, and perform chemistry. Cryo-electron microscopy (cryo-EM) and cryo-electron tomography (cryo-ET) can access the intrinsic heterogeneity of these complexes and are therefore key tools for understanding their function. However, 3D reconstruction of the collected imaging data presents a challenging computational problem, especially without any starting information, a setting termed ab initio reconstruction. Here, in this study, we introduce cryoDRGN-AI, a method leveraging an expressive neural representation and combining an exhaustive search strategy with gradient-based optimization to process challenging heterogeneous datasets. Using cryoDRGN-AI, we reveal new conformational states in large datasets, reconstruct previously unresolved motions from unfiltered datasets, and demonstrate ab initio reconstruction of biomolecular complexes from in situ data. With this expressive and scalable model for structure determination, we hope to unlock the full potential of cryo-EM and cryo-ET as a high-throughput tool for structural biology and discovery.

Levy, Axel [Stanford Univ., CA (United States); SL↗

DOE OSTI2025

Gradient flow based phase-field modeling using separable neural networks

Allen–Cahn equation is a reaction–diffusion equation and is widely used for modeling phase separation. Machine learning methods for solving the Allen–Cahn equation in its strong form suffer from inaccuracies in collocation techniques, errors in computing higher-order spatial derivatives, and the large system size required by the space–time approach. To overcome these challenges, we propose solving the gradient flow of the Ginzburg–Landau free energy functional, which is equivalent to the Allen–Cahn equation, thereby avoiding the second-order spatial derivatives associated with the Allen–Cahn equation. A minimizing movement scheme is employed to solve the gradient flow problem, eliminating the complexities of a space–time approach. We utilize a separable neural network that efficiently represents the phase field through low-rank tensor decomposition. As we use the minimizing movement scheme to numerically solve the gradient flow problem, we thus, refer to the proposed method as the Separable Deep Minimizing Movement (SDMM) method. The evaluation of the functional in the minimizing movement scheme using the Gauss quadrature technique bypasses the inaccuracies associated with collocation techniques traditionally used to solve partial differential equations. A hyperbolic tangent transformation is introduced on the phase field prior to the evaluation of the functional to ensure that it remains strictly bounded within the values of the two phases. For this transformation, theoretical guarantee for energy stability of the minimizing movement scheme is established. Our results suggest that this transformation helps to improve the accuracy and efficiency significantly. The proposed method resolves the challenges faced by state-of-the-art machine learning techniques, outperforming them in both accuracy and efficiency. It is also the first machine learning method to achieve an order of magnitude speed improvement over the finite element method. In addition to its formulation and computational implementation, several case studies illustrate the applicability of the proposed method.

42 ENGINEERING↗

DOE OSTI2024

On learning what to learn: Heterogeneous observations of dynamics and establishing possibly causal relations among them

Abstract Before we attempt to (approximately) learn a function between two sets of observables of a physical process, we must first decide what the inputs and outputs of the desired function are going to be. Here we demonstrate two distinct, data-driven ways of first deciding “the right quantities” to relate through such a function, and then proceeding to learn it. This is accomplished by first processing simultaneous heterogeneous data streams (ensembles of time series) from observations of a physical system: records of multiple observation processes of the system. We determine (i) what subsets of observables are common between the observation processes (and therefore observable from each other, relatable through a function); and (ii) what information is unrelated to these common observables, therefore particular to each observation process, and not contributing to the desired function. Any data-driven technique can subsequently be used to learn the input–output relation—from k-nearest neighbors and Geometric Harmonics to Gaussian Processes and Neural Networks. Two particular “twists” of the approach are discussed. The first has to do with the identifiability of particular quantities of interest from the measurements. We now construct mappings from a single set of observations from one process to entire level sets of measurements of the second process, consistent with this single set. The second attempts to relate our framework to a form of causality: if one of the observation processes measures “now,” while the second observation process measures “in the future,” the function to be learned among what is common across observation processes constitutes a dynamical model for the system evolution.

Sroczynski, David W.↗

DOE OSTI2026

Gearbox bearing crack growth prognostics and uncertainty quantification with physics-informed machine learning

This paper introduces the extreme theory of functional connections (X-TFC), a physics-informed machine learning algorithm, and tailors it to estimate the remaining useful life (RUL) of wind turbine gearbox bearings experiencing fatigue crack growth. Unlike purely data-driven methods, X-TFC embeds a physics model, based on Head's theory in this work, into its training objective. The core of X-TFC is a random-projection single-layer neural network trained via an extreme learning machine, which requires only limited damage progression data and solves for output weights with a least-squares optimization algorithm. A composite loss function balances the network's fit to observed degradation data against the residuals of the governing crack growth differential equation, ensuring the learned damage trajectory remains physically plausible. When applied to a vibration-based health-index (HI) dataset measured during the growth of a crack on the inner ring of a high-speed bearing in a wind turbine gearbox (Bechhoefer and Dubé, 2020), X-TFC achieves near-zero prediction bias. Even when trained on only the first 10 %–20 % of the damage progression data, with sufficient physics weighting its predictions remain monotonic and smooth, delivering high prognosability and trendability. To quantify the epistemic uncertainty, we employ a Monte Carlo ensemble of independently initialized X-TFC models trained on noise-perturbed data, which yields confidence intervals around each RUL estimate and captures both model-parameter and epistemic uncertainty. In addition to a vibration-based HI, we demonstrate that the proposed framework can be directly applied to a supervisory control and data acquisition (SCADA) data-based HI (Eftekhari Milani et al., 2026) measured during similar wind turbine gearbox bearing crack faults, preserving its accuracy and interpretability. This extension shows the versatility of our approach, which is applicable to bearings of multiple gearbox manufacturers, models, and ratings using only SCADA data. By integrating domain knowledge with machine learning, X-TFC offers a rapid, reliable tool for crack prognostics. Its adaptability to other bearing failure modes, such as pitch bearing ring cracks, positions X-TFC as a powerful enabler of data-driven, physics-informed asset management in the wind energy sector and beyond.

17 WIND ENERGY↗

DOE OSTI2024

Exploring Li-Ion Transport Properties of Li 3 TiCl 6 : A Machine Learning Molecular Dynamics Study

We performed large-scale molecular dynamics simulations based on a machine-learning force field (MLFF) to investigate the Li-ion transport mechanism in cation-disordered Li 3 TiCl 6 cathode at six different temperatures, ranging from 25°C to 100°C. In this work, deep neural network method and data generated by ab − initio molecular dynamics (AIMD) simulations were deployed to build a high-fidelity MLFF. Radial distribution functions, Li-ion mean square displacements (MSD), diffusion coefficients, ionic conductivity, activation energy, and crystallographic direction-dependent migration barriers were calculated and compared with corresponding AIMD and experimental data to benchmark the accuracy of the MLFF. From MSD analysis, we captured both the self and distinct parts of Li-ion dynamics. The latter reveals that the Li-ions are involved in anti-correlation motion that was rarely reported for solid-state materials. Similarly, the self and distinct parts of Li-ion dynamics were used to determine Haven’s ratio to describe the Li-ion transport mechanism in Li 3 TiCl 6 . Obtained trajectory from molecular dynamics infers that the Li-ion transportation is mainly through interstitial hopping which was confirmed by intra- and inter-layer Li-ion displacement with respect to simulation time. Ionic conductivity (1.06 mS/cm) and activation energy (0.29eV) calculated by our simulation are highly comparable with that of experimental values. Overall, the combination of machine-learning methods and AIMD simulations explains the intricate electrochemical properties of the Li 3 TiCl 6 cathode with remarkably reduced computational time. Thus, our work strongly suggests that the deep neural network-based MLFF could be a promising method for large-scale complex materials.

Selvaraj, Selva Chandrasekaran (ORCID:000000029023↗

DOE OSTI2024

Quantifying local and global mass balance errors in physics-informed neural networks

Physics-informed neural networks (PINN) have recently become attractive for solving partial differential equations (PDEs) that describe physics laws. By including PDE-based loss functions, physics laws such as mass balance are enforced softly in PINN. This paper investigates how mass balance constraints are satisfied when PINN is used to solve the resulting PDEs. We investigate PINN’s ability to solve the 1D saturated groundwater flow equations (diffusion equations) for homogeneous and heterogeneous media and evaluate the local and global mass balance errors. We compare the obtained PINN’s solution and associated mass balance errors against a two-point finite volume numerical method and the corresponding analytical solution. We also evaluate the accuracy of PINN in solving the 1D saturated groundwater flow equation with and without incorporating hydraulic heads as training data. We demonstrate that PINN’s local and global mass balance errors are significant compared to the finite volume approach. Tuning the PINN’s hyperparameters, such as the number of collocation points, training data, hidden layers, nodes, epochs, and learning rate, did not improve the solution accuracy or the mass balance errors compared to the finite volume solution. Mass balance errors could considerably challenge the utility of PINN in applications where ensuring compliance with physical and mathematical properties is crucial.

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2024

Photometric redshifts probability density estimation from recurrent neural networks in the DECam local volume exploration survey data release 2

Photometric wide-field surveys are imaging the sky in unprecedented detail. These surveys face a significant challenge in efficiently estimating galactic photometric redshifts while accurately quantifying associated uncertainties. In this work, we address this challenge by exploring the estimation of Probability Density Functions (PDFs) for the photometric redshifts of galaxies across a vast area of 17,000 square degrees, encompassing objects with a median 5 σ point-source depth of g = 24.3, r = 23 . 9 , i = 23.5, and z = 22.8 mag. Our approach uses deep learning, specifically integrating a Recurrent Neural Network architecture with a Mixture Density Network, to leverage magnitudes and colors as input features for constructing photometric redshift PDFs across the whole DECam Local Volume Exploration (DELVE) survey sky footprint. Subsequently, we rigorously evaluate the reliability and robustness of our estimation methodology, gauging its performance against other well-established machine learning methods to ensure the quality of our redshift estimations. Our best results constrain photometric redshifts with the bias of − 0 . 0013 , a scatter of 0.0293, and an outlier fraction of 5.1%. These point estimates are accompanied by well-calibrated PDFs evaluated using diagnostic tools such as Probability Integral Transform and Odds distribution. We also address the problem of the accessibility of PDFs in terms of disk space storage and the time demand required to generate their corresponding parameters.We present a novel Autoencoder model that reduces the size of PDF parameter arrays to one-sixth of their original length, significantly decreasing the time required for PDF generation to one-eighth of the time needed when generating PDFs directly from the magnitudes.

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2025

Physics-informed latent neural operator for real-time predictions of time-dependent parametric PDEs

Deep operator network (DeepONet) has shown significant promise as surrogate models for systems governed by partial differential equations (PDEs), enabling accurate mappings between infinite-dimensional function spaces. However, when applied to systems with high-dimensional input-output mappings arising from large numbers of spatial and temporal collocation points, these models often require heavily overparameterized networks, leading to long training times. Latent DeepONet addresses some of these challenges by introducing a two-step approach: first learning a reduced latent space using a separate model, followed by operator learning within this latent space. While efficient, this method is inherently data-driven and lacks mechanisms for incorporating physical laws, limiting its robustness and generalizability in data-scarce settings. Here, in this work, we propose PI-Latent-NO, a physics-informed latent neural operator framework that integrates governing physics directly into the learning process. Our architecture features two coupled DeepONets trained end-to-end: a Latent-DeepONet that learns a low-dimensional representation of the solution, and a Reconstruction-DeepONet that maps this latent representation back to the physical space. By embedding PDE constraints into the training via automatic differentiation, our method eliminates the need for labeled training data and ensures physics-consistent predictions. The proposed framework is both memory and compute-efficient, exhibiting near-constant scaling with problem size and demonstrating significant speedups over traditional physics-informed operator models. We validate our approach on a range of parametric PDEs, showcasing its accuracy, scalability, and suitability for real-time prediction in complex physical systems.

Latent representations↗

DOE OSTI2025

Generalizable machine learning potentials for quantum-accurate predictions of non-equilibrium behavior in 2D materials

Machine learning interatomic potentials (ML-IAPs) are emerging as transformative tools in materials modeling, promising quantum-level accuracy at a fraction of the computational cost. However, their ability to generalize beyond equilibrium configurations and to reliably capture defect- and temperature-driven behavior remains underexplored. Here, we develop and benchmark two state-of-the-art ML-IAPs, Spectral Neighbor Analysis Potential (SNAP) and Allegro, on a comprehensive dataset for monolayer MoSe₂. Using density functional theory (DFT) as the reference, we evaluate their performance in capturing stress–strain behavior, phase transition energetics, defect evolution, edge stability, and fracture toughness. Allegro, a deep equivariant neural network potential, surpasses both SNAP and the classical Tersoff potential in accuracy, efficiency, and transferability. Importantly, both ML potentials accurately reproduce experimental fracture measurements and ab initio predictions of inversion domain formation—phenomena well beyond their training sets. Our findings establish ML-IAPs as viable replacements for traditional force fields in the study of non-equilibrium mechanical phenomena, enabling large-scale, high-fidelity simulations in 2D materials and beyond. In conclusion, this work provides a broadly applicable framework for data-driven modeling of structural and functional transformations under extreme conditions.

2D materials↗

DOE OSTI2024

Challenges in Training PINNs: A Loss Landscape Perspective

This paper explores challenges in training Physics Informed Neural Networks (PINNs), emphasizing the role of the loss landscape in the training process. We examine difficulties in minimizing the PINN loss function, particularly due to ill conditioning caused by differential operators in the residual term. We compare gradient-based optimizers Adam, L-BFGS, and their combination Adam+L-FGS, showing the superiority of Adam+L-BFGS, and introduce a novel secondorder optimizer, NysNewton-CG (NNCG), which significantly improves PINN performance. Theoretically, our work elucidates the connection between ill-conditioned differential operators and ill-conditioning in the PINN loss and shows the benefits of combining first- and second-order optimization methods. Our work presents valuable insights and more powerful optimization strategies for training PINNs, which could improve the utility of PINNs for solving difficult partial differential equations.

Rathore, Pratik↗

DOE OSTI2025

Introduction: Neuromorphic Materials

The explosive growth in data collection and the need to process it efficiently, as well as the desire to automate increasingly complex tasks in transportation, medical care, manufacturing, security and many other fields have motivated a growing interest in neuromorphic computing. Unlike the binary, transistorbased ON/OFF logic gates and separate logic and memory functionalities employed in digital computing, neuromorphic computing is inspired by animal brains that use interconnected synapses and neurons to perform processing, storage and transmission of information at the same location, while only consuming ~20 W or less of power. Motivated by the brain’s efficiency, adaptability, self-learning and resiliency qualities, neuromorphic computing can be broadly defined as an approach to processing and storing information using hardware and algorithms inspired by models of biological neural systems. Present research in neuromorphic computing encompasses approaches that vary significantly in their degree of neuro-inspiration, from systems that only incorporate features such as asynchronous, event-driven operation or use crossbar arrays of non-volatile memory (NVM) elements to accelerate deep neural networks (DNNs), to designs that embrace the extreme parallelism, sparsity, reconfigurability, adaptability, complexity and stochasticity observed in nervous systems. The term ‘neuromorphic’ computing is often credited to Carver Mead, who in the 1980s investigated Si-based analog electronics to replicate functions of the animal retina. Earlier important advances in this field include the work of Frank Rosenblatt, who proposed the concept of the perceptron, Bernard Widrow, who used this concept to build one of the first analog neural networks, the Adaline and many other researchers (see ref. 6 for an historical perspective on neuromorphic computing). With the recent increase in the use of artificial intelligence and large language models, and rising concerns over the associated energy costs, interest in neuromorphic hardware has expanded rapidly. According to some estimates, driven largely by the drastic growth in the training use of artificial intelligence (AI) models using the current computing architectures, the energy cost of computing is projected to reach the energy supply worldwide by 2045. Furthermore, while this is not a realistic outcome, it means that, if more efficient computing technologies are not developed -- soon -- the world will soon become one where demand for energy and market constraints limit the continued increase of societal access to AI and cloud services from data centers. Data centers used for training and use of these models consume hundreds of terawatt hours of electricity, already past 4% of the US electricity demand.

Circuits↗

DOE OSTI2026

TEMPEST

This repository solves the problem of driver identification through vehicular and biometric data. Through an embedding-based approach and a novel loss function, we're able to distinguish between different drivers' behaviors. This also provides preprocessing for reproducibility of results.The code preprocesses vehicular data, trains neural networks, and outputs predictions.This code introduces a novel embedding-based neural network with a 91% rank-1 accuracy, as well as all code to reproduce training and results.

Musgrove, Kyle↗

DOE OSTI2025

CHEMREASONER: Heuristic Search over a Large Language Model’s Knowledge Space using Quantum-Chemical Feedback

The discovery of new catalysts is essential for the design of new and more efficient chemical processes in order to transition to a sustainable future. We introduce an AI-guided computational screening framework unifying linguistic reasoning with quantum-chemistry based feedback from 3D atomistic representations. Our approach formulates catalyst discovery as an uncertain environment where an agent actively searches for highly effective catalysts via the iterative combination of large language model (LLM)-derived hypotheses and atomistic graph neural network (GNN)-derived feedback. Identified catalysts in intermediate search steps undergo structural evaluation based on spatial orientation, reaction pathways, and stability. Scoring functions based on adsorption energies and barriers steer the exploration in the LLM's knowledge space toward energetically favorable, high-efficiency catalysts. We introduce planning methods that automatically guide the exploration without human input, providing competitive performance against expert-enumerated chemical descriptor-based implementations. By integrating language-guided reasoning with computational chemistry feedback, our work pioneers AI-accelerated, trustworthy catalyst discovery.

artificial intelligence↗

DOE OSTI2025

Safe Physics-Informed Machine Learning for Dynamics and Control

This tutorial paper focuses on safe physics-informed machine learning in the context of dynamics and control, providing a comprehensive overview of how to integrate physical models and safety guarantees. As machine learning techniques enhance the modeling and control of complex dynamical systems, ensuring safety and stability remains a critical challenge, especially in safety-critical applications like autonomous vehicles, robotics, medical decision-making, and energy systems. We explore various approaches for embedding and ensuring safety constraints, including structural priors, Lyapunov and Control Barrier Functions, predictive control, projections, and robust optimization techniques. Additionally, we delve into methods for uncertainty quantification and safety verification, including reachability analysis and neural network verification tools, which help validate that control policies remain within safe operating bounds even in uncertain environments. The paper includes illustrative examples demonstrating the implementation aspects of safe learning frameworks that combine the strengths of data-driven approaches with the rigor of physical principles, offering a path toward the safe control of complex dynamical systems.

Drgona, Jan↗

DOE OSTI2026

Peak2Patch: High-Fidelity Functional Group Identification through Attention-Based Fusion of Infrared and Mass Spectra

Identifying molecular structure based on spectroscopic readings is a key task in a variety of chemical and biological applications. Common spectroscopy techniques, such as Infrared (IR) Spectroscopy and Mass Spectrometry (MS), provide detailed information on the structure of molecular compounds but nonetheless require expert-level knowledge to decode. Machine learning has emerged as a potential solution for automating structure prediction from chemical spectra; however, current approaches generally focus on single sensor modalities, neglecting to leverage the complementary information contained within differing spectra. In this paper, we introduce Peak2Patch, a novel approach to fusion-enhanced prediction of functional groups from IR and mass spectra. First, we perform a detailed comparison of backbone networks for encoding both sparse mass spectra and dense IR spectra and demonstrate the superior performance of transformer neural networks over current state-of-the-art convolutional neural networks. Second, we evaluate three broad categories of fusion: early (raw feature), middle (deep feature), and late (decision) fusion, demonstrating the potential of a deep feature fusion-based approach. Lastly, we present Peak2Patch, our attention-based fusion scheme, which leverages cross-attention to mix features between encoded tokens of the two modalities. We validate our approach on a publicly available multimodal spectroscopic data set of 790k simulated molecules, demonstrating a large improvement in functional group prediction over both the previous state-of-the-art and our own strong single-modal baselines.

Jacobson, Philip [Sandia National Laboratories (SN↗